Personalized Federated Learning with Multi-Server

In recent years, multi-server federated learning has gained significant attention in cross-device scenarios due to its capability to reduce communication delay. Classic multi-server federated learning algorithms simply use clients located in overlapping service areas to aggregate the model parameters from regional servers they can access, which ultimately train a common global model. However, the global model obtained through this approach often exhibit poor generalization performance on heterogeneous data. To overcome this challenge, we propose personalized federated learning based on adaptive model fusion (pFedAF). This method employs a Local Pre-aggregation (LP) module to adaptively aggregate the local models of clients with regional models, ultimately generating personalized models suited to the clients' private data. Experimental results demonstrate that the proposed pFedAF can effectively improve model performance on non-independent and identically distributed (Non-IID) data. Compared to other multi-server federated learning algorithms, our pFedAF algorithm demonstrates superior performance in Non-IID scenarios.